CN-121980508-A - Industrial process fault diagnosis method based on intervention enhanced invariant causal graph learning network

CN121980508ACN 121980508 ACN121980508 ACN 121980508ACN-121980508-A

Abstract

The invention provides an industrial process fault diagnosis method based on an intervention enhanced invariance causal graph learning network, which comprises the steps of 1, preprocessing process data of process variables collected in a target industrial operation process, constructing a variable association graph reflecting topological relation among the variables, 2, establishing an invariance causal graph learning network, 3, performing intervention enhancement, constructing an intervention distribution environment, 4, synthesizing a causal sub-graph prediction result and a false sub-graph prediction result based on a joint prediction mechanism, introducing invariance pre-risks by using causal invariance, identifying a stable fault causal sub-graph structure under different intervention distribution environments, determining causal action of the fault causal sub-graph structure on fault categories, describing a fault mechanism and a propagation path, and realizing fault diagnosis according to the fault causal sub-graph structure. Compared with other methods, the fault diagnosis method provided by the invention can improve the reliability and the interpretability of the fault diagnosis of the industrial process.

Inventors

ZHANG XIANGRUI
GE LIN
DAI WEI
WANG LANHAO
LIU XIN
NAN JING

Assignees

中国矿业大学

Dates

Publication Date: 20260505
Application Date: 20260126

Claims (10)

1. An industrial process fault diagnosis method based on an intervention enhanced invariant causal graph learning network is characterized by comprising the following steps: Step 1, for a target industrial process, carrying out data preprocessing on process data of process variables acquired in the operation process to obtain standardized time sequence data, generating dynamic time sequence data through a fixed-length sliding window, and constructing a variable association graph reflecting topological relation among the variables based on a Pearson correlation coefficient; step 2, establishing a constant causal graph learning network IICGLN, wherein the constant causal graph learning network IICGLN comprises a causal generator, a graph encoder and a classifier, dividing a variable associated graph input causal graph into a fault causal graph and an environment false graph, and extracting the characteristics of the divided fault causal graph and environment false graph by using a shared graph encoder respectively; Step 3, performing intervention enhancement based on the fault causal subgraph and the environment false subgraph divided by the causal generator, and actively constructing an intervention distribution environment by recombining the fault causal subgraph of the current training sample and the environment false subgraph of other training samples; And 4, based on a combined prediction mechanism, synthesizing a causal sub-graph prediction result and a false sub-graph prediction result, utilizing causal invariance to introduce a non-drying pre-risk, identifying a stable fault causal sub-graph structure under different intervention distribution environments, and finally, based on an industrial process fault diagnosis model of an intervention enhanced invariable causal graph learning network, only using pure causal prediction to actively exclude the influence of the environment false sub-graph identified in training, so as to ensure that a fault diagnosis result is completely determined by the stable fault causal sub-graph.
2. The method of claim 1, wherein in step 1, the data preprocessing comprises: The representation comprises Training data of individual samples, wherein Representing the T-th sample containing n process variables, T representing the transpose, Representing real space, robust normalized data preprocessing of process data of the process variable, and the first Time series of individual variables , Wherein Represent the first Nth sample value of the variable, using the nth Median of individual variables Deviation from median Scaling to obtain standardized data Wherein Median represents a median function; Definition of the definition Wherein Represents the normalized data value of the nth variable in the nth sample, Processing the normalized time series data using a sliding window, the input matrix of the t th window being Wherein Indicating the size of a sliding window, wherein the fault label corresponding to the window is that 。
3. The method of claim 2 wherein in step 1, a variable association graph is created from standardized data within a sliding window, a single process variable is considered as a node in the graph, and a set of nodes is defined Wherein Representing the nth process variable, the ith process variable being in the current window The normalized sample sequence in the network is node By pearson correlation coefficient Quantization node And Correlation strength at the current time window: , Wherein the method comprises the steps of Representing nodes Sum node Covariance within the current window, Representing nodes The standard deviation within the current window is used, Representing nodes The standard deviation within the current window is used, Representing variables The first in the current window The data of the plurality of data, Representing variables The mean value within the current window; based on preset threshold value Screening the association strength, only reserving strong association higher than a threshold value as an edge, and constructing a sparse adjacency matrix : , Wherein the method comprises the steps of Elements representing the ith row and jth column of the sparse adjacency matrix a; If and only if the correlation coefficient When there is an edge in the graph Constructing an undirected graph The correlation structure between the variables within the window is represented as input to a subsequent fault diagnosis model based on the invariance causal graph learning network IICGLN.
4. A method as claimed in claim 3, wherein in step 2 the cause and effect generator is to generate an undirected graph Dividing into fault causal subgraphs And context spurious subgraphs Firstly, extracting node characteristics of the variable association graph Mask matrix for computing edges based on node characteristics , Representing nodes Sum node Importance scores between, expressed as: , , Wherein, the And Respectively represent the passes through Extracted node Feature representation vectors and nodes of (a) Is a feature representation vector of (a), Representing the Sigmoid activation function, The graph feature extraction network is formed by stacking local extremum graph convolution layers; Based on mask matrix And the original adjacency matrix Determining importance scores of each edge, and measuring contribution degrees of the edges to fault diagnosis prediction; selecting the edge with the highest score of the front r proportion to form a fault causal subgraph according to a preset causal dividing proportion r Edge set of (2) Selecting the rest edges with low scores as environment false subgraphs Edge set of (2) And find out two sets of edges And Corresponding nodes respectively construct fault causal subgraphs And context spurious subgraphs Expressed as: , , Wherein, the The matrix elements are arranged according to the importance scores of all sides, the sides corresponding to the front r proportion with the largest numerical value are reserved, Refers to element-by-element product.
5. The method of claim 4, wherein in step 2, the causal subgraph is constructed for faults in an intervention environment And context spurious subgraphs Coding and independent classification are carried out, a model is forced to capture a cross-environment stable fault causal subgraph structure, and a shared graph encoder is utilized first Respectively fault causal subgraph Coding with an environmental false subgraph s, generating a fault causal subgraph using GNN coding Node characteristic representation of (c) And node feature representation of an environmental spurious subgraph s Node characteristics are then represented using global pooling Aggregation into graph-level feature representations Representing node characteristics Aggregation into graph-level feature representations Expressed as: , , Wherein, the Is convolved by local extremum map ) The layer-stacked graph feature extraction network Pooling refers to a global pooling operation; Predicting a fault causal subgraph c and an environment false subgraph s respectively by using two independent classifiers, wherein the causal classifier Graph level feature representation based on fault causal subgraph Causal prediction is performed, and prediction of causal parts of output faults is performed Learning a root cause and effect mechanism of the fault; False classifier Graph level representation based on environmental spurious subgraphs Making a false prediction, outputting a prediction of a pure false portion The predictive power of the false correlation of the environment on the fault label is measured.
6. The method of claim 5, wherein in step 3, an intervention-enhanced training process is performed to actively construct more than two intervention distribution environments by reorganizing the fault causal subgraph of the current training sample with the environment false subgraphs of other training samples, first dividing the environment false subgraphs after causal generation of all fault variable association graphs g Collecting, constructing environment false sub-graph library, executing intervention operation Wherein Representing the ambient false sub-graph variables, Representing a sample of a selected concrete environmental false sub-graph instance, by combining the current sample Is a causal subgraph of the fault of (a) Fixing and replacing the original environment false sub-graph part with the environment false sub-graph of other samples in the library Creating a new intervention sample 。
7. The method of claim 6, wherein in step 4, the model incorporates causal predictions during training And false predictions And (3) carrying out joint prediction: , Wherein the confidence level Representing the reliability degree of the environment false subgraph on fault diagnosis, if the environment false correlation is reliable Representing confidence level Trend 1, joint prediction retains causal subgraph predictors ; If the environmental dummy subgraph is not effective for fault diagnosis Representing confidence level Trending toward 0, resulting in a joint prediction result Near 0.
8. The method of claim 7, wherein in step 4, there is no risk of drying out The calculation formula is as follows: , Wherein, the Indicating that the actual failure signature is one, A fault diagnosis model is represented and is used for the diagnosis of faults, Representing the weight of the variance term in the balance super parameter used for adjusting the non-drying pre-risk, different distribution environments constructed under intervention, obtaining a group of losses under different intervention environments, And Representing the mean and variance of the risk values respectively, Representing a risk loss function, and calculating the formula as follows: , Wherein the method comprises the steps of Representing a cross-entropy loss function, Calculating expectations under a particular intervention profile, at the non-intervention data profile In, maintain fault causal structure Causal structure for original sample The environment false structure is an environment false subgraph under specific intervention ; Causal predictive loss The calculation formula is as follows: , Wherein, the Representing expectations on the raw data distribution; Training a false classifier with independent environmental false loss functions Identifying false environment structure, and during training, Updating only false classifiers Parameters of (c) environmental spurious loss function The calculation formula is as follows: , Total loss function The method comprises the following steps: , The model after training is only used for purely causal prediction when fault diagnosis is carried out, and firstly an input graph g is passed through a causal generator Extracting a fault causal subgraph c using only encoders and causal classifiers Is based on a divided fault causal subgraph To make final fault causal predictions Discarding false classifier by actively rejecting interference of environmental false subgraphs identified in training And a joint prediction mechanism, ensuring that the fault diagnosis decision is entirely determined by the stable fault causal subgraph.
9. An electronic device comprising a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method of any of claims 1 to 8.
10. A storage medium storing a computer program or instructions which, when run on a computer, performs the steps of the method of any one of claims 1 to 8.

Description

Industrial process fault diagnosis method based on intervention enhanced invariant causal graph learning network Technical Field The invention belongs to the field of process monitoring and fault diagnosis, and particularly relates to an industrial process fault diagnosis method based on an intervention enhanced invariant causal graph learning network. Background Modern industrial processes are bulky in structure and complex in internal mechanism of production process, and occurrence of faults often causes chain reactions to cause significant loss and safety risks. Therefore, the industrial process fault diagnosis has important significance for guaranteeing the safety of industrial production and improving the production efficiency. There is a large amount of monitored variable data in actual industrial processes, and there is a complex correlation between these variables. Faults in industrial processes do not occur independently, but rather propagate along a fault path causing abnormal changes in a number of related variables. The Graph Neural Network (GNN) is capable of processing graph structure data using the neural network, and the GNN's messaging mechanism causes each node to aggregate information of its neighboring nodes to update the central node state. The fault diagnosis model established based on the GNN can enhance the accuracy of fault diagnosis and is excellent in graph classification tasks in multiple fields. Modeling the fault diagnosis of the industrial process as a graph classification task, wherein each fault corresponds to a different graph structure, and realizing the fault diagnosis by identifying and classifying the graph structures of different faults. The existing industrial process fault diagnosis method based on the GNN mostly describes the correlation among variables by establishing a correlation graph among the variables, and the poor generalization capability of the method leads to the performance reduction of fault diagnosis. However, most of the existing industrial process causal discovery studies convert multivariate sensor data into graph structures based on causal relationships defined by prediction, such as the methods of granj causal analysis, transfer entropy and the like. However, most of these approaches stay at the level of statistical relevance, neglecting the mining and utilization of fault causal subgraph structures that are causally decisive for fault categories and that can describe the mechanism and propagation paths of the fault. The goal of existing methods is typically to find statistical associations between input graphs and true fault labels, however such associations do not necessarily reflect true causal relationships, and the established causal graph contains a large number of false correlation associations. Therefore, the model is easily influenced by shortcut characteristics derived from factors such as sample selection or environmental noise in training data when fault diagnosis is decided, the shortcut characteristics do not have cross-environment stability, so that the model has weak generalization capability outside distribution, and the performance is sharply reduced when the working condition is changed or new noise appears. Meanwhile, the decision dependence is not causal characteristic in the data, so that the process of fault diagnosis is weak in interpretation. In an industrial process, it is difficult to trace the propagation path of occurrence of a fault and the cause of occurrence of the fault. Therefore, a method is adopted to excavate a fault cause and effect sub-graph structure from the fault variable association graph structure, so that the phenomenon that a shortcut structure is learned is avoided, the cause and effect structure which truly determines the fault type is learned, the fault diagnosis is carried out by utilizing the fault cause and effect sub-graph structure, and the generalization capability, stability and interpretability of an industrial process fault diagnosis model are improved. Disclosure of Invention Aiming at the defects of the prior art, the invention provides an industrial process fault diagnosis method based on an intervention enhanced invariant causal graph learning network, which comprises the following steps: Step 1, for a target industrial process, carrying out data preprocessing on process data of process variables acquired in the operation process to obtain standardized time sequence data, generating dynamic time sequence data through a fixed-length sliding window, and constructing a variable association graph reflecting topological relation among the variables based on a Pearson correlation coefficient; step 2, establishing a constant causal graph learning network IICGLN, wherein the constant causal graph learning network IICGLN comprises a causal generator, a graph encoder and a classifier, dividing a variable associated graph input causal graph into a fault causal graph and an environment false graph,